hashmap

Overview

A Golang lock-free thread-safe HashMap optimized for fastest read access.

It is not a general-use HashMap and currently has slow write performance for write heavy uses.

The minimal supported Golang version is 1.19 as it makes use of Generics and the new atomic package helpers.

Usage

Example uint8 key map uses:

m := New[uint8, int]()
m.Set(1, 123)
value, ok := m.Get(1)

Example string key map uses:

m := New[string, int]()
m.Set("amount", 123)
value, ok := m.Get("amount")

Using the map to count URL requests:

m := New[string, *int64]()
var i int64
counter, _ := m.GetOrInsert("api/123", &i)
atomic.AddInt64(counter, 1) // increase counter
...
count := atomic.LoadInt64(counter) // read counter

Benchmarks

Reading from the hash map for numeric key types in a thread-safe way is faster than reading from a standard Golang map in an unsafe way and four times faster than Golang's sync.Map:

ReadHashMapUint-8                676ns ± 0%
ReadHaxMapUint-8                 689ns ± 1%
ReadGoMapUintUnsafe-8            792ns ± 0%
ReadXsyncMapUint-8               954ns ± 0%
ReadGoSyncMapUint-8             2.62µs ± 1%
ReadSkipMapUint-8               3.27µs ±10%
ReadGoMapUintMutex-8            29.6µs ± 2%

Reading from the map while writes are happening:

ReadHashMapWithWritesUint-8      860ns ± 1%
ReadHaxMapWithWritesUint-8       930ns ± 1%
ReadGoSyncMapWithWritesUint-8   3.06µs ± 2%

Write performance without any concurrent reads:

WriteGoMapMutexUint-8           14.8µs ± 2%
WriteHashMapUint-8              22.3µs ± 1%
WriteGoSyncMapUint-8            69.3µs ± 0%

The benchmarks were run with Golang 1.19.1 on Linux and a Ryzen 9 5900X CPU using make benchmark-perflock.

Technical details

• Technical design decisions have been made based on benchmarks that are stored in an external repository: go-benchmark

• The library uses a sorted linked list and a slice as an index into that list.

• The Get() function contains helper functions that have been inlined manually until the Golang compiler will inline them automatically.

• It optimizes the slice access by circumventing the Golang size check when reading from the slice. Once a slice is allocated, the size of it does not change. The library limits the index into the slice, therefore the Golang size check is obsolete. When the slice reaches a defined fill rate, a bigger slice is allocated and all keys are recalculated and transferred into the new slice.

• For hashing, specialized xxhash implementations are used that match the size of the key type where available